No backward in the feature extraction network

[kloudkl]

Because the learning rate is default to be 1 and backward is default to be true (#389), even though the phase is set to Caffe::TEST, the feature extraction network still performs backward propagation unnecessarily.

[sguada]

Although lr=1 the network only perform backward propagation if you call to the backward function. So although the log says the layer needs backward propagation it will not do it at least it is requested.

[Yangqing]

I second Sergio's comment - it is not necessary to force it. No backward propagation is carried out unless you call backward. There is actually a good research reason to keep non-zero learning rate (albeit a little hacky) - one can examine gradients if s/he wants and that sometimes reveals salient regions in the images.

[Noiredd]

If backward propagation is not explicitly called, are the gradients still being calculated during inference? I'm optimizing a network for inference speed, and for this particular application I'm not interested in backprop - caffe time still includes it in measurements though. Ideally, I'd like to be able to train a network and then disable all unnecessary computations in order to benchmark it. Do I have to modify the entire proto, or is there some known best way to approach this?